Download Document

NOVAMENTE
A Practical Architecture for Artificial General Intelligence
Ben Goertzel, PhD
Novamente LLC
Biomind LLC
Artificial General Intelligence Research Institute
Virginia Tech, Applied Research Lab for National and Homeland Security
The Novamente Project
•
Long-term goal:
–
•
Novamente AI Engine: an integrative AI architecture
–
–
–
•
creating "artificial general intelligence" approaching and then exceeding the
human level
Overall design founded on a unique holistic theory of intelligence
Cognition carried out via computer science algorithms rather than imitation of
human brain
efficient, scalable C++/Linux implementation
Currently, isolated parts of the Novamente codebase are being used
for commercial projects
–
–
natural language processing
biological data analysis
Overview Papers
• The Novamente AI Engine
– IJCAI Workshop on Intelligent Control of Agents, Acapulco,
August 2003
• Novamente: An Integrative Architecture for Artificial General
Intelligence
– AAAI Symposium on Achieving Human-Level Intelligence
Through Integrated Systems and Research, Washington DC,
October 2004
• Patterns, Hypergraphs and General Intelligence
– World Congress on Computational Intelligence, Vancover CA,
July 2006
• Chapter on Novamente in
– Artificial General Intelligence volume, Springer Verlag, 2006
This edited volume -- the first ever to focus
exclusively on Artificial General Intelligence -- is
edited by Dr. Ben Goertzel and Cassio Pennachin
and contains chapters by AGI researchers at
universities, corporations and research institutes
around the world.
A partial author list includes:
- Ben Goertzel (Novamente LLC)
- Cassio Pennachin (Novamente LLC)
- Marcus Hutter (IDSIA)
- Juergen Schmidhuber (ISDIA)
- Pei Wang (Temple University)
- Peter Voss (A2I2)
- Vladimir Redko (Keldysh Institute)
- Eliezer Yudkowsky (SIAI)
- Lukasz Kaiser (Aachen Univ. of Technology)
Novamente AI Engine
Components of the system have
been commercially deployed
– Biomind OnDemand product for
bioinformatic data analysis
– ImmPort: NIH Web portal with
Biomind/Novamente based
analytics on the back end
– INLINK language processing
system developed for INSCOM
(Army Intelligence)
The Grand Vision
–
–
–
–
–
–
Conceptual Background
Teaching Approach
Knowledge Representation
Software Architecture
Cognitive Processes
Emergent Mental Structures
The Current Reality
– Implemented Components
– Simulation-World Experiments
The Path Ahead
Novamente:
The Grand Vision
Conceptual Background:
Patternist Philosophy of Mind
• An intelligent system is conceived as a system for
recognizing patterns in the world and in itself
• Probability theory may be used as a language for
quantifying and relating patterns
• Logic (term, predicate, combinatory) may be used as
a base-level language for expressing patterns
• The reflexive process of flexibly recognizing patterns
in oneself and then improving oneself based on these
patterns is the “basic algorithm of intelligence”
• The phenomenal self, a key aspect of intelligent
systems, is the result of an intelligent system
recognizing itself as a pattern in its (internal and
external) behaviors
Conceptual Background:
Definition of Intelligence
• Intelligence is considered as the ability to
achieve complex goals in a complex
environment
• Goals are achieved via recognizing
probabilistic patterns of the form “Carrying out
procedure P in context C will achieve goal G.”
Patternist Philosophy

Minds are systems of
patterns that achieve goals
by recognizing patterns in
themselves and the world

AI is about creating
software whose structures
and dynamics will lead to
the emergence of these
pattern-sets
Prior, Conceptually Relevant
Book Publications
The Structure of Intelligence, Springer-Verlag, 1993
The Evolving Mind, Gordon and Breach, 2003
Chaotic Logic, Plenum Press, 1994
From Complexity to Creativity, Plenum Press, 1997
Creating Internet Intelligence, Kluwer Academic, 2001
Novamente-Related Books-in-Progress
• Probabilistic
Term Logic
–In final editing stage; to be submitted 2006
• Engineering General Intelligence
–In final editing stage
–Reviews the overall NM design
–May or may not be submitted for publication (AI Safety and
commmercial concerns)
• Artificial Cognitive Development
–Developmental psychology for Novamente and other AGIs
–In preparation
AI Teaching Methodology
• Embodiment
• Post-embodiment
• Developmental Stages
Embodiment in AGISim
Simulation World
Post-Embodied AI
AI systems may viably synthesize knowledge
gained via various means
•
virtually embodied experience
– AGISim
•
physically embodied experience
– Robotics
•
explicit encoding of knowledge
– in natural language
– In artificial languages such as Lojban, Lojban++
•
ingestion of databases
– WordNet, FrameNet, Cyc, etc.
– quantitative scientific data
Stages of Cognitive Development
Knowledge Representation
Novamente’s “Atom Space”
• Atoms = Nodes or Links
• Atoms have
– Truth values (probability + weight of evidence)
– Attention values (short and long term importance)
• The Atomspace is a weighted, labeled
hypergraph
Novamente’s “Atom Space”
• Not a neural net
– No activation values, no attempt at low-level brain modeling
– But, Novamente Nodes do have “attention values”, analogous to
time-averages of neural net activations
• Not a semantic net
– Atoms may represent percepts, procedures, or parts of concepts
– Most Novamente Atoms have no corresponding English label
– But, most Novamente Atoms do have probabilistic truth values,
allowing logical semantics
Attention Values
Low Long-term Importance
High Long-term Importance
Low Short-term
Importance
Useless
Remembered but not
currently used (e.g.
mother’s phone #)
High Short-term
Importance
Used then forgotten
(e.g. most precepts)
Used and remembered
Truth Values
Strength low
Strength high
Weight of
evidence low
Weakly suspected to be
false
Weakly suspected to be
true
Weight of
evidence high
Firmly known to be false
Firmly known to be true
Atoms Come in Various Types
• ConceptNodes
•
– “tokens” for links to attach
to
• PredicateNodes
• ProcedureNodes
• PerceptNodes
– Visual, acoustic percepts,
etc.
• NumberNodes
Logical links
–
–
–
–
–
•
•
InheritanceLink
SimilarityLink
ImplicationLink
EquivalenceLink
Intensional logical relationships
HebbianLinks
Procedure evaluation links
Links may denote generic association …
…or precisely specified relationships
Software Architecture &
Cognitive Architecture
Simplified Workflow
Feelings
Goals
Execution
Management
Percepts
Active
Memory
World
Active Schema
Pool
Cognitive Processes
Typology of Cognitive Processes
Global processes
• MindAgents that
periodically iterate
through all Atoms and act
on them
• “Things that all Atoms do”
Control Processes
• Execution of actions
• Maintenance of goal
hierarchy
• Updating of system control
schemata
Focused processes
• MindAgents that begin by
selecting a small set of
important or relevant Atoms,
and then act on these to
generate a few more small
sets of Atoms, and iterate
• Two species:
– Forward synthesis
– Backward synthesis
Global Cognitive Processes
• Attention Allocation
– Updates short and long term importance values associated
with Atoms
– Uses a “simulated economy” approach, with separate
currencies for short and long term importance
• Stochastic pattern mining of the AtomTable
– A powerful heuristic for predicate formation
– Critical for perceptual pattern recognition as well as
cognition
– Pattern mining of inference histories critical to advanced
inference control
• Building of the SystemActivityTable
– Records which MindAgents acted on which Atoms at which
times
– Table is used for building HebbianLinks, which are used in
attention allocation
Control Processes
• Execution of procedures
– “Programming language interpreter” for
executable procedures created from NM Atoms
• Maintenance of “active procedure pool”
– Set of procedures that are currently ready to be
activated if their input conditions are met
• Maintenance of “active goal pool”
– Set of predicates that are currently actively
considered as system goals
Global Cognitive Processes, Part I
Forward Synthesis
Forward Synthesis Processes
• Forward-Chaining Probabilistic Inference
– Given a set of knowledge items, figure out what
(definitely or speculatively) follows from it
• Concept/Goal Formation
– “Blend” existing concepts or goals to form new
ones
• Map formation
– Create new Atoms out of sets of Atoms that tend
to be simultaneously important (or whose
importance tends to be coordinated according to
some other temporal pattern)
Forward Synthesis Processes
• Language Generation
– Atoms representing semantic relationships are
combined with Atoms representing linguistic
mapping rules to produce Atoms representing
syntactic relationships, which are then
transformed into sentences
• Importance Propagation
– Atoms pass some of their “attentional currency” to
Atoms that they estimate may help them become
important again in the future
“Probabilistic Logic Networks” (PLN) for uncertain inference
Example First-Order PLN Rules Acting on ExtensionalInheritanceLinks
A  B
B  C
|A C
A
B
C
Deduc tion
A  B
A  C
|B  C
A
B
C
Induction
A  C
B  C
|A  B
A
B
C
Abduction
Grounding of natural language constructs is provided via inferential
integration of data gathered from linguistic and perceptual inputs
Novamente contains multiple heuristics for Atom creation, including
“blending” of existing Atoms
Atoms associated in a dynamic “map” may be grouped to form new
Atoms: the Atomspace hence explicitly representing patterns in itself
Global Cognitive Processes, Part II
Backward Synthesis
Backward Synthesis Processes
• Backward-chaining probabilistic inference
– Given a target Atom, find ways to produce and evaluate it
logically from other knowledge
• Inference process adaptation
– Given a set of inferential conclusions, find ways to produce
those conclusions more effectively than was done before
• Predicate Schematization
– Given a goal, and knowledge about how to achieve the goal,
synthesize a procedure for achieving the goal
• Credit Assignment
– Given a goal, figure out which procedures’ execution, and
which Atoms’ importance, can be expected to lead to the
goal’s achievement
• Goal Refinement
– Given a goal, find other (sub)goals that imply that goal
Insert A-not-B screenshot
(Partial) PLN Backward-Chaining Inference
Trajectory for Piagetan A-not-B Problem
Step 3
Modus Ponens
Target:
Eval found_under(toy_6,$1)
Step 1
ANDRule:
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
|AND <1.00, 0.98>
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
Step 2
Unification:
Imp <1.00, 0.94>
AND
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
Eval found_under(toy_6, red_bucket_6)
AND <1.00, 0.98>
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
|Eval found_under(toy_6, red_bucket_6) <1.00, 0.93>
Imp <1.00, 0.95>
AND
Inh($t,toy)
Inh($b,bucket)
Eval placed_under($t,$b)
Eval found_under($t,$b)
AND
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
|Imp <1.00, 0.94>
AND
Inh (toy_6,toy)
Inh (red_bucket_6,bucket)
Eval placed_under(toy_6,red_bucket_6)
Eval found_under(toy_6,red_bucket_6)
The system may study its own inference history to figure out inference control
patterns that would have let it arrive at its existing knowledge more effectively.
This is a type of backward synthesis that may lead to powerful iterative selfimprovement.
Predicate Schematization
Logical knowledge
Executable procedure
EvPredImp <0.95, 0.3>
Execution try(goto box)
Eval near box
SimultaneousImplication
Eval near box
Eval can_do(push box)
EvPredImp <0.6,0.4>
And
Eval can_do(push box)
Execution try(push box)
Evaluation Reward
Predicate schematization
repeat
goto box
near box
repeat
push box
Reward
(More)
Backward Synthesis Processes
• Model-Based Predicate Generation
– Given probabilistic knowledge about what patterns
characterize predicates or procedures satisfying a
certain criterion, generate new
predicate/procedures satisfying the criterion
• Criterion-Based Predicate Modeling
– Building of probabilistic knowledge regarding the
patterns characterizing predicates satisfying a
certain criterion
As shown in Moshe Looks’ PhD thesis work, the combination of the above two
processes may play the role of evolutionary programming, but with dramatically
better performance on many problem cases, and an enhanced capability to carry
out learning across multiple fitness functions (criteria).
MOSES: Meta-Optimizing Semantic Evolutionary Search
Bringing evolutionary programming and probabilistic inference together
•
MOSES evolved out of BOA Programming, which was an extension to program tree learning
of the Bayesian Optimization Algorithm approach to probabilistic evolutionary learning
•
May be fully integrated with PLN backward chaining inference as a special kind of
“backward synthesis process”
–
•
Algorithm:
–
–
–
–
•
Integration currently incomplete, to be completed in 2007
a population of procedure/predicate trees are evaluated
the best ones are simplified and normalized …
… and modeled probabilistically (Criterion-Based Predicate Modeling)
Then new trees are generated via instance generation based on these probabilistic models (ModelBased Predicate Generation)
Moshe Looks PhD Thesis 2006, Washington University, St. Louis
–
www.metacog.org
Simple Example: A MOSES Population of
Arithmetic Procedures
Simplification &
Normalization
Before Normalization
Normalization of
procedure/predicate
trees harmonizes
syntactic form with
semantic meaning (I/O
behavior)
After Normalization
Syntactic Distance
Graphs based on
Boolean predicates;
same phenomenon
holds more generally
Syntactic Distance
Alignment
(Recognizing common patterns)
Abstract trees (predicates) are created from the
population of concrete ones
ifelse
holding
ifelse
Example: MOSES learns
program to play “fetch”
in AGISim
facingteacher
step
rotate
ifelse
nearball
pickup
ifelse
facingball
step
rotate
(More)
Backward Synthesis Processes
Language Comprehension
– Syntax parsing: given a sentence, or other utterance,
search for assignments of syntactic relationships to words
that will make the sentence grammatical
– Semantic mapping: Search for assignment of semantic
meanings to words and syntactic relationships that will make
the sentence contextually meaningful
Lojban / Lojban++
Lojban is a constructed language
with syntax and semantics
founded on predicate logic
Lojban++ is a variant of Lojban
that incorporates English
content words in certain roles
In these languages, ambiguity is
minimized relative to natural
languages
Parsing Lojban/++ is automatic
and mechanical
Semantic mapping into predicate
logic is also fully mechanical -but some contextual
disambiguation of predicates
may still be required
Lojban / Lojban++
English
Lojban
I eat the salad with croutons
mi citka le salta poi mixre lo sudnabybli
Lojban++ mi eat le salad poi mixre lo crouton
mi eat le salad poi contain lo crouton
English
I eat the salad with a fork
Lojban
mi citka le salta sepi'o lo forca
Lojban++ mi eat le salad sepi'o lo fork
Lojban++
le dog pe mi uncle cu stupid
EvaluationLink stupid $D
InheritanceLink $D dog
Needs contextual
disambiguation
AssociationLink $D $U
EvaluationLink uncle($U, Ben_Goertzel)
Holistic Cognitive Dynamics
and Emergent Mental
Structures
The Fundamental Cognitive Dynamic
Let X = any set of Atoms
Let F(X) = a set of Atoms which is the result of forward synthesis on
X
Let B(X) = a set of Atoms which is the result of backward synthesis
of X -- assuming a heuristic biasing the synthesis process
toward simple constructs
Let S(t) denote a set of Atoms at time t, representing part of a
system’s knowledge base
Let I(t) denote Atoms resulting from the external environment at
time t
S(t+1) = B( F(S(t) + I(t)) )
The Fundamental Cognitive Dynamic
S(t+1) = B( F(S(t) + I(t)) )
Forward: create new mental forms by combining
existing ones
Backward: seek simple explanations for the forms in the
mind, including the newly created ones. The
explanation itself then comprises additional new
forms in the mind
Forward: …
Backward: …
Etc.
… Combine … Explain … Combine … Explain … Combine …
The Construction and Development
of the Emergent Pattern that is the
“Phenomenal Self”
The self originates (and
ongoingly re-originates)
via backward synthesis
Backward chaining
inference attempts to
find models that will
explain the observed
properties of the system
itself
The self develops via forward
synthesis
Aspects of self blend with each
other and combine
inferentially to form new
Atoms
These new Atoms help guide
behavior, and thus become
incorporated into the
backward-synthesis-derived
self-models
Self = A strange attractor of the Fundamental Cognitive Dynamic
The Construction and Development
of the Emergent Pattern that is
“Focused Consciousness”
Atoms in the “moving
bubble of importance”
consisting of the Atoms
with highest Short-Term
Importance are
continually combining
with each other, forming
new Atoms that in many
cases remain highly
important
Sets of Atoms in the moving
bubble of importance are
continually subjected to
backward synthesis, leading
to the creation of compact
sets of Atoms that
explain/produce them -- and
these new Atom-sets often
remain highly important
Focused Consciousness = A strange attractor of the
Fundamental Cognitive Dynamic
Why Will Novamente Succeed Where
Other AGI Approaches Fail?
• Only Novamente is based on
a well-reasoned, truly
comprehensive theory of
mind, covering both the
concretely-implemented and
emergent aspects
• The specific algorithms and
data structures chosen to
implement this theory of
mind are efficient, robust and
scalable
• So is the software
implementation!
More specifically: Only in
the Novamente design is
the fundamental cognitive
dynamic implemented in a
powerful and general
enough way adequate to
give rise to self and focused
consciousness as strange
attractors.
Novamente:
The Current Reality
Implemented Components
•
•
•
•
•
•
Novamente core system
– AtomTable, MindAgents, Scheduler, etc.
– Now runs on one machine; designed for distributed
processing
PLN
– Relatively crude inference control heuristics
– Simplistic predicate schematization
MOSES
– Little experimentation has been done evolving
procedures with complex control structures
– Not yet fully integrated with PLN
Schema execution framework
– Enacts learned procedures
AGISim
– And proxy for communication with NM core
NLP front end
– External NLP system for “cheating” style knowledge
ingestion
Simple, Initial
AGISim Experiments
•
•
•
•
Fetch
Tag
Piagetan A-not-B experiment
Word-object association
Goal For Year One After Project Funding
Fully Functional Artificial Infant
Able to learn infant-level behaviors "without
cheating" -- i.e. with the only instruction being
interactions with a human-controlled agent in the
simulation world
Example behaviors: naming objects, asking for
objects, fetching objects, finding hidden objects,
playing tag
System will be tested using a set of tasks
derived from human developmental psychology
Within first 9 months after funding we plan to
have the most capable autonomous artificial
intelligent agent created thus far, interacting with
humans spontaneously in its 3D simulation
world in the manner of a human infant
Teaching the Baby Language
Artificial Infant + Narrow-AI NLP System =
AGI system capable of learning complex natural
language
(Narrow-AI NLP system as “scaffolding”)
Narrow-AI NLP System =
Novamente’s RelEx English semantic analysis
engine + a Lojban++ parser
(Parallel instruction in English and Lojban++ will accelerate
learning dramatically)
Goal For Year Two After Project Funding
Artificial Child with Significant
Linguistic Ability
Ability to learn from human teachers
via linguistic communication utilizing
complex recursive phrase structure
grammar and grounded semantics
Linguistic instruction will be done
simultaneously in English and in the
constructed language Lojban++, which
maps directly into formal logic
At this stage, the symbol groundings
learned by the system will be
commercially very valuable, and will
be able to dramatically enhance the
performance of natural language
question answering products
Acknowledgements
The Novamente Team
•
Bruce Klein – President, Novamente LLC
•
Cassio Pennachin – Chief Architect,
Novamente AI Engine
•
Andre Senna – CTO
•
Ari Heljakka – Lead AI Engineer
•
Moshe Looks – AI Engineer
•
Izabela Goertzel– AI Engineer
Dr. Matthew Ikle’
Bruce Klein
Dr. Moshe Looks
Ari Heljakka
•
Murilo Queiroz – AI Engineer
•
Welter Silva – System Architect
•
Dr. Matthew Ikle’ – Mathematician
Dr. Ben Goertzel
Izabela Goertzel
2006 AGIRI.org Workshop
Sponsored by Novamente LLC)
Cassio Pennachin
Thank You